performance_tests: add stats and loop count multiplier options

Stats are: min, median, standard deviation

1 files changed, 104 insertions, 19 deletions

diff --git a/tests/performance_tests/performance_tests.h b/tests/performance_tests/performance_tests.h
index ce9baf61f..d37dda729 100644
--- a/tests/performance_tests/performance_tests.h
+++ b/tests/performance_tests/performance_tests.h
@@ -36,6 +36,9 @@
 #include <boost/chrono.hpp>
 #include <boost/regex.hpp>
 
+#include "misc_language.h"
+#include "common/perf_timer.h"
+
 class performance_timer
 {
 public:
@@ -62,14 +65,21 @@ private:
   clock::time_point m_start;
 };
 
+struct Params
+{
+  bool verbose;
+  bool stats;
+  unsigned loop_multiplier;
+};
 
 template <typename T>
 class test_runner
 {
 public:
-  test_runner(bool verbose = true)
+  test_runner(const Params &params)
     : m_elapsed(0)
-    , m_verbose(verbose)
+    , m_params(params)
+    , m_per_call_timers(T::loop_count * params.loop_multiplier, {true})
   {
   }
 
@@ -82,14 +92,18 @@ public:
     performance_timer timer;
     timer.start();
     warm_up();
-    if (m_verbose)
+    if (m_params.verbose)
       std::cout << "Warm up: " << timer.elapsed_ms() << " ms" << std::endl;
 
     timer.start();
-    for (size_t i = 0; i < T::loop_count; ++i)
+    for (size_t i = 0; i < T::loop_count * m_params.loop_multiplier; ++i)
     {
+      if (m_params.stats)
+        m_per_call_timers[i].resume();
       if (!test.test())
         return false;
+      if (m_params.stats)
+        m_per_call_timers[i].pause();
     }
     m_elapsed = timer.elapsed_ms();
 
@@ -101,9 +115,62 @@ public:
   int time_per_call(int scale = 1) const
   {
     static_assert(0 < T::loop_count, "T::loop_count must be greater than 0");
-    return m_elapsed * scale / T::loop_count;
+    return m_elapsed * scale / (T::loop_count * m_params.loop_multiplier);
+  }
+
+  uint64_t per_call_min() const
+  {
+    uint64_t v = std::numeric_limits<uint64_t>::max();
+    for (const auto &pt: m_per_call_timers)
+      v = std::min(v, pt.value());
+    return v;
+  }
+
+  uint64_t per_call_max() const
+  {
+    uint64_t v = std::numeric_limits<uint64_t>::min();
+    for (const auto &pt: m_per_call_timers)
+      v = std::max(v, pt.value());
+    return v;
+  }
+
+  uint64_t per_call_mean() const
+  {
+    uint64_t v = 0;
+    for (const auto &pt: m_per_call_timers)
+      v += pt.value();
+    return v / m_per_call_timers.size();
   }
 
+  uint64_t per_call_median() const
+  {
+    std::vector<uint64_t> values;
+    values.reserve(m_per_call_timers.size());
+    for (const auto &pt: m_per_call_timers)
+      values.push_back(pt.value());
+    return epee::misc_utils::median(values);
+  }
+
+  uint64_t per_call_stddev() const
+  {
+    if (m_per_call_timers.size() <= 1)
+      return 0;
+    const uint64_t mean = per_call_mean();
+    uint64_t acc = 0;
+    for (const auto &pt: m_per_call_timers)
+    {
+      int64_t dv = pt.value() - mean;
+      acc += dv * dv;
+    }
+    acc /= m_per_call_timers.size () - 1;
+    return sqrt(acc);
+  }
+
+  uint64_t min_time_ns() const { return tools::ticks_to_ns(per_call_min()); }
+  uint64_t max_time_ns() const { return tools::ticks_to_ns(per_call_max()); }
+  uint64_t median_time_ns() const { return tools::ticks_to_ns(per_call_median()); }
+  uint64_t standard_deviation_time_ns() const { return tools::ticks_to_ns(per_call_stddev()); }
+
 private:
   /**
    * Warm up processor core, enabling turbo boost, etc.
@@ -122,30 +189,39 @@ private:
 private:
   volatile uint64_t m_warm_up;  ///<! This field is intended for preclude compiler optimizations
   int m_elapsed;
-  bool m_verbose;
+  Params m_params;
+  std::vector<tools::PerformanceTimer> m_per_call_timers;
 };
 
 template <typename T>
-void run_test(const std::string &filter, bool verbose, const char* test_name)
+void run_test(const std::string &filter, const Params &params, const char* test_name)
 {
   boost::smatch match;
   if (!filter.empty() && !boost::regex_match(std::string(test_name), match, boost::regex(filter)))
     return;
 
-  test_runner<T> runner(verbose);
+  test_runner<T> runner(params);
   if (runner.run())
   {
-    if (verbose)
+    if (params.verbose)
     {
       std::cout << test_name << " - OK:\n";
-      std::cout << "  loop count:    " << T::loop_count << '\n';
+      std::cout << "  loop count:    " << T::loop_count * params.loop_multiplier << '\n';
       std::cout << "  elapsed:       " << runner.elapsed_time() << " ms\n";
+      if (params.stats)
+      {
+        std::cout << "  min:       " << runner.min_time_ns() << " ns\n";
+        std::cout << "  max:       " << runner.max_time_ns() << " ns\n";
+        std::cout << "  median:    " << runner.median_time_ns() << " ns\n";
+        std::cout << "  std dev:   " << runner.standard_deviation_time_ns() << " ns\n";
+      }
     }
     else
     {
-      std::cout << test_name << " (" << T::loop_count << " calls) - OK:";
+      std::cout << test_name << " (" << T::loop_count * params.loop_multiplier << " calls) - OK:";
     }
     const char *unit = "ms";
+    uint64_t scale = 1000000;
     int time_per_call = runner.time_per_call();
     if (time_per_call < 30000) {
      time_per_call = runner.time_per_call(1000);
@@ -154,8 +230,17 @@ void run_test(const std::string &filter, bool verbose, const char* test_name)
 #else
      unit = "µs";
 #endif
+     scale = 1000;
+    }
+    std::cout << (params.verbose ? "  time per call: " : " ") << time_per_call << " " << unit << "/call" << (params.verbose ? "\n" : "");
+    if (params.stats)
+    {
+      uint64_t min_ns = runner.min_time_ns() / scale;
+      uint64_t med_ns = runner.median_time_ns() / scale;
+      uint64_t stddev_ns = runner.standard_deviation_time_ns() / scale;
+      std::cout << " (min " << min_ns << " " << unit << ", median " << med_ns << " " << unit << ", std dev " << stddev_ns << " " << unit << ")";
     }
-    std::cout << (verbose ? "  time per call: " : " ") << time_per_call << " " << unit << "/call" << (verbose ? "\n" : "") << std::endl;
+    std::cout << std::endl;
   }
   else
   {
@@ -164,10 +249,10 @@ void run_test(const std::string &filter, bool verbose, const char* test_name)
 }
 
 #define QUOTEME(x) #x
-#define TEST_PERFORMANCE0(filter, verbose, test_class)         run_test< test_class >(filter, verbose, QUOTEME(test_class))
-#define TEST_PERFORMANCE1(filter, verbose, test_class, a0)     run_test< test_class<a0> >(filter, verbose, QUOTEME(test_class<a0>))
-#define TEST_PERFORMANCE2(filter, verbose, test_class, a0, a1) run_test< test_class<a0, a1> >(filter, verbose, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ">")
-#define TEST_PERFORMANCE3(filter, verbose, test_class, a0, a1, a2) run_test< test_class<a0, a1, a2> >(filter, verbose, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ">")
-#define TEST_PERFORMANCE4(filter, verbose, test_class, a0, a1, a2, a3) run_test< test_class<a0, a1, a2, a3> >(filter, verbose, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ">")
-#define TEST_PERFORMANCE5(filter, verbose, test_class, a0, a1, a2, a3, a4) run_test< test_class<a0, a1, a2, a3, a4> >(filter, verbose, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ", " QUOTEME(a4) ">")
-#define TEST_PERFORMANCE6(filter, verbose, test_class, a0, a1, a2, a3, a4, a5) run_test< test_class<a0, a1, a2, a3, a4, a5> >(filter, verbose, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ", " QUOTEME(a4) ", " QUOTEME(a5) ">")
+#define TEST_PERFORMANCE0(filter, params, test_class)         run_test< test_class >(filter, params, QUOTEME(test_class))
+#define TEST_PERFORMANCE1(filter, params, test_class, a0)     run_test< test_class<a0> >(filter, params, QUOTEME(test_class<a0>))
+#define TEST_PERFORMANCE2(filter, params, test_class, a0, a1) run_test< test_class<a0, a1> >(filter, params, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ">")
+#define TEST_PERFORMANCE3(filter, params, test_class, a0, a1, a2) run_test< test_class<a0, a1, a2> >(filter, params, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ">")
+#define TEST_PERFORMANCE4(filter, params, test_class, a0, a1, a2, a3) run_test< test_class<a0, a1, a2, a3> >(filter, params, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ">")
+#define TEST_PERFORMANCE5(filter, params, test_class, a0, a1, a2, a3, a4) run_test< test_class<a0, a1, a2, a3, a4> >(filter, params, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ", " QUOTEME(a4) ">")
+#define TEST_PERFORMANCE6(filter, params, test_class, a0, a1, a2, a3, a4, a5) run_test< test_class<a0, a1, a2, a3, a4, a5> >(filter, params, QUOTEME(test_class) "<" QUOTEME(a0) ", " QUOTEME(a1) ", " QUOTEME(a2) ", " QUOTEME(a3) ", " QUOTEME(a4) ", " QUOTEME(a5) ">")